Prize: an R package for prioritization estimation based on analytic hierarchy process

Transcription

1 Prize: an R package for prioritization estimation based on analytic hierarchy process Daryanaz Dargahi October 30, 2017 daryanazdargahi@gmail.com Contents 1 Licensing 2 2 Overview 2 3 Relative AHP Defining the problem and determining the criteria, subcriteria, and alternatives Structuring the decision hierarchy Constructing pairwise comparison matrices Aggregating individual judgments into a group judgment Estimating and visualizing priorities Rating AHP Defining a rating scale and obtaining alternatives priorities

2 1 Licensing Under the Artistic License, you are free to use and redistribute this software. 2 Overview The high throughput studies often produce large amounts of numerous genes and proteins of interest. While it is difficult to study and validate all of them. In order to narrow down such lists, one approach is to use a series of criteria to rank and prioritize the potential candidates based on how well they meet the research goal. Analytic hierarchy process (AHP) [1] is one of the most popular group decision-making techniques for ranking and prioritizing alternatives when multiple criteria must be considered. It provides a comprehensive and rational framework to address complicated decisions by modeling the problem in a hierarchical structure, showing the relationships of the goal, objectives (criteria and subcriteria), and alternatives. AHP has unique advantages when communication among team members is impeded by their different specializations or perspectives. It also enables decision makers to evaluate decision alternatives when important elements of the decision are difficult to quantify or compare. The AHP technique uses pairwise comparisons to measure the impact of items on one level of the hierarchy on the next higher level. It has two models for arriving at a ranking of alternatives. (A) The relative model, where alternatives are compared in a pairwise manner regarding their ability to achieve each of the criteria. (B) The rating model is often used when the number of alternatives is large, or if the possibility of adding or deleting alternatives exists [2]. This model requires establishing a series of rating scales (categories) for each criterion. These scales must be pairwise compared to determine the relative importance of each rating category, and then alternatives are evaluated one at a time by selecting the appropriate rating category for each criterion. Here, we introduce an R package for AHP, Prize. Prize offers the implementation of both relative and rating AHP models. In order to rank and prioritize a set of alternatives with AHP, decision makers must take four steps: 1. Define the problem and determine the criteria, subcriteria, and alternatives 2. Structure the decision hierarchy 3. Construct pairwise comparison matrices 4. Estimate and visualize priorities In the following, we describe a brief example use case for Prize in translational oncology. 2

3 3 Relative AHP 3.1 Defining the problem and determining the criteria, subcriteria, and alternatives Assume a scenario that a group of scientists identified 10 genes that are being differentially expressed (DE) in tumor tissues in comparison to healthy tissues. They are interested in ranking and prioritizing these genes based on their potential role as a tumor marker or therapeutic target. They decide to consider the (1) gene expression profile in tumor tissue, (2) gene expression profile in healthy tissue, (3) frequency of being DE, and (4) epitopes as the criteria for making their decision. They also subdivide the epitope criterion into the size and number of extracellular regions. 3.2 Structuring the decision hierarchy The scientists form their decision hierarchy as follows; > require(prize) > require(diagram) > mat <- matrix(nrow = 7, ncol = 2, data = NA) > mat[,1] <- c('0', '1','2','3','4','4.1','4.2') > mat[,2] <- c('prioritization_of_de_genes','tumor_expression','normal_expression', + 'Frequency', 'Epitopes', 'Number_of_epitopes', 'Size_of_epitopes') > mat [,1] [,2] [1,] "0" "Prioritization_of_DE_genes" [2,] "1" "Tumor_expression" [3,] "2" "Normal_expression" [4,] "3" "Frequency" [5,] "4" "Epitopes" [6,] "4.1" "Number_of_epitopes" [7,] "4.2" "Size_of_epitopes" > ahplot(mat, fontsize = 0.7, cradx = 0.11,sradx = 0.12, cirx= 0.18, ciry = 0.07) 3

4 Prioritization_of_DE_genes Tumor_expression Normal_expression Frequency Epitopes Number_of_epitopes Size_of_epitopes 3.3 Constructing pairwise comparison matrices Each scientist (decision maker) investigates the values of the decision elements in the hierarchy, and incorporates their judgments by performing a pairwise comparison of these elements. Each decision element in the upper level is used to compare the elements of an immediate inferior level of the hierarchy with respect to the former. That is, the alternatives are compared with respect to the subcriteria, the subcriteria are compared with respect to the criteria and the criteria are compared with respect to the goal. Therefore, each decision maker constructs a set of pairwise comparison matrices reflecting how important decision elements are to them with respect to the goal. Pairwise comparison matrices are built from the comparison between elements, based on the Saaty fundamental scale [3]. 4

5 Table 1: Saaty fundamental scale for pairwise comparison [3] Intensity of Definition Explanation importance 1 Equal importance Two elements contribute equally to the objective 3 Moderate importance Experience and judgement slightly favor one element over an other 5 Strong importance Experience and judgement strongly favor one element over an other 7 Very strong importance 9 Extreme importance One element is favored very strongly over an other, its dominance is demonstrated in practice The evidence favoring one element over another is of the highest possible order of affirmation Intensities of 2,4,6, and 8 can be used to express intermediate values. Intensitise 1.1, 1.2, 1.3, etc. can be used for elements that are very close in importance For instance, to pairwise compare the criteria a total of six comparisons must be done, including Tumor expression/normal expression, Tumor expression/frequency, Tumor expression/epitope, Normal expression/frequency, Normal expression/epitope, and Frequency/Epitope. The criteria pairwise comparison matrix (PCM) is shown below. > pcm <- read.table(system.file('extdata','ind1.tsv',package = 'Prize'), + sep = '\t', header = TRUE, row.names = 1) > pcm Tumor_expression Normal_expression Frequency Epitopes Tumor_expression Normal_expression NA Frequency NA NA 1 2 Epitopes NA NA NA 1 The ahmatrix function completes a pairwise comparison matrix by converting the triangular matrix into a square matrix, where diagonal values are equal 1 and pcm[j,i] = 1/pcm[i,j]. 5

6 > pcm <- ahmatrix(pcm) > ahp_matrix(pcm) Tumor_expression Normal_expression Frequency Epitopes Tumor_expression Normal_expression Frequency Epitopes Aggregating individual judgments into a group judgment Once the individual PCMs are available, gaggregate function could be used to combine the opinions of various decision makers into an overall opinion for the group. gaggregate offers two aggregation methods including aggregation of individual judgments (AIJ - geometric mean) and aggregation of individual priorities (AIP - using arithmetic mean) [4]. If decision makers have different expertise or perspectives, in order to reflect that in the group judgment, one can use a weighted AIJ or AIP, by simply providing a weight for each decision maker. > mat = matrix(nrow = 4, ncol = 1, data = NA) > mat[,1] = c(system.file('extdata','ind1.tsv',package = 'Prize'), + system.file('extdata','ind2.tsv',package = 'Prize'), + system.file('extdata','ind3.tsv',package = 'Prize'), + system.file('extdata','ind4.tsv',package = 'Prize')) > rownames(mat) = c('ind1','ind2','ind3', 'ind4') > colnames(mat) = c('individual_judgement') > # non-weighted AIJ > res = gaggregate(srcfile = mat, method = 'geometric', simulation = 500) > # aggregated group judgement using non-weighted AIJ > AIJ(res) Tumor_expression Normal_expression Frequency Epitopes Tumor_expression Normal_expression Frequency Epitopes > # consistency ratio of the aggregated group judgement > GCR(res) [1] The distance among individual and group judgments can be visualized using the dplot function. dplot uses a classical multidimensional scaling (MDS) approach [5] to compute the distance among individual and group priorities. 6

7 > require(ggplot2) > # Distance between individual opinions and the aggregated group judgement > dplot(ip(res)) 0.03 ind ind Coordinate Group judgement ind1 ind Coordinate 1 The consistency ratio of individual judgments can be visualized using the crplot function. If the consistency ratio is equal or smaller than 0.1, then the decision is considered to be consistent. > # Consistency ratio of individal opinions > crplot(icr(res), angle = 45) 7

8 ICR Condition Failed Pass ind1 ind2 ind3 ind4 ID 3.4 Estimating and visualizing priorities In order to obtain the priorities of decision elements to generate the final alternatives priorities, local and global priorities are required to be obtained from the comparison matrices. Local priorities are determined by computing the maximum eigenvalue of the PCMs. The local priorities are then used to ponder the priorities of the immediately lower level for each element. The global priorities are obtained by multiplying the local priorities of the elements by the global priority of their above element. The total priorities of the alternatives are found by the addition of alternatives global prioritiese. The pipeline function computes local and global priorities, as well as final prioritization values. P ipeline can simply be called by a matrix including the 8

9 problem hierarchy and group PCMs. The scientists use the following matrix (mat) to call the pipeline function; > require(stringr) > mat <- matrix(nrow = 7, ncol = 3, data = NA) > mat[,1] <- c('0', '1','2','3','4','4.1','4.2') > mat[,2] <- c('prioritization_of_de_genes','tumor_expression','normal_expression', + 'Frequency', 'Epitopes', 'Number_of_epitopes', 'Size_of_epitopes') > mat[,3] <- c(system.file('extdata','aggreg.judgement.tsv',package = 'Prize'), + system.file('extdata','tumor.pcm.tsv',package = 'Prize'), + system.file('extdata','normal.pcm.tsv',package = 'Prize'), + system.file('extdata','freq.pcm.tsv',package = 'Prize'), + system.file('extdata','epitope.pcm.tsv',package = 'Prize'), + system.file('extdata','epitopenum.pcm.tsv',package = 'Prize'), + system.file('extdata','epitopelength.pcm.tsv',package = 'Prize')) > # Computing alternatives priorities > prioritization <- pipeline(mat, model = 'relative', simulation = 500) The global priorities of decision elements can be visualized using the ahplot function. > ahplot(ahp_plot(prioritization), fontsize = 0.7, cradx = 0.11,sradx = 0.12, + cirx= 0.18, ciry = 0.07, dist = 0.06) 9

10 Prioritization_of_DE_genes Tumor_expression Normal_expression Frequency Epitopes Number_of_epitopes Size_of_epitopes Contribution of decision elements in the final priority estimation could also be visualized using wplot. > require(reshape2) > wplot(weight_plot(prioritization)$criteria_wplot, type = 'pie', + fontsize = 7, pcex = 3) 10

11 12% 7% Criteria Tumor_expression 47% Normal_expression 34% Frequency Epitopes The rainbow function illustrates prioritized alternatives detailing the contribution of each criterion in the final priority score. > rainbowplot(rainbow_plot(prioritization)$criteria_rainbowplot, xcex = 3) 11

12 FZD CD CD MUC Alternative EGFR 1956 CD Criteria Tumor_expression Normal_expression Frequency Epitopes BSG CD MUC CA Total priority score The Carbonic anhydrase 9 (CA9) and Mucin-16 (MUC16) with a global priority of are the alternative that contribute the most to the goal of choosing the optimal tumor marker/therapeutic target among the identified DE genes. Drugs targeting CA9 and MUC16 are currently in pre-clinical and clinical studies [6, 7]. > rainbow_plot(prioritization)$criteria_rainbowplot Tumor_expression Normal_expression Frequency Epitopes BSG CD CD CA MUC CD

13 EGFR CD MUC FZD total_priorities BSG CD CD CA MUC CD EGFR CD MUC FZD Rating AHP As the number of alternatives increase, the amount of pairwise comparison becomes large. Therefore, pairwise comparisons take much time and also the possibility of inconsistency in the comparisons increases. Rating AHP overcomes this problem by categorizing the criteria and/or subcriteria in order to classify alternatives. In another words, rating AHP uses a set of categories that serves as a base to evaluate the performance of the alternatives in terms of each criterion and/or subcriterion. The rating procedure is also suitable when the possibility of adding/removing alternatives exists. The rating AHP reduces the number of judgments that decision makers are required to make. The rating AHP differs from the relative AHP in the evaluation and obtaining the priority of alternatives. Hence, the decision markers define their decision problem, structure the problem into a hierarchy, and collect PCM matrices for each criteria/subcriteria similar to the relative AHP approach. Then, they use a rating approach to evaluate alternatives. 4.1 Defining a rating scale and obtaining alternatives priorities In the example scenario, the scientists would like to rank and prioritize 10 genes based on their potential role as a tumor marker/therapeutic target. To build a PCM matrix consisting of 10 alternatives 45 pairwise comparisons are required. The large number of pairwise comparisons makes this step time consuming and increase the possibility of inconsistency in the comparisons. Therefore, scientists decide to use rating AHP by defining a series of categories with respect to the criteria and/or subcriteria to evaluate alternatives. They also compute a PCM of these categories. For instance, they define two categories, single and multiple, for the numberof epitopes subcriteria, and compute their PCM. 13

14 > category_pcm = read.table(system.file('extdata','number.tsv', package = 'Prize') +, sep = '\t', header = TRUE, row.names = 1) > category_pcm Single Multiple Single 1 2 Multiple NA 1 Then, decision makers evaluate the alternatives against the defined categories and build an alternative matrix showing the category that each alternative belongs to. > alt_mat = read.table(system.file('extdata','numepitope_alternative_category.tsv', + package = 'Prize'), sep = '\t', header = FALSE) > alt_mat V1 V2 1 BSG 682 Single 2 CD Single 3 CD Single 4 CA9 768 Single 5 MUC Single 6 CD9 928 Multiple 7 EGFR 1956 Single 8 CD Single 9 MUC Single 10 FZD Multiple To compute the idealised priorities of alternatives, the rating functions can be called by a category PCM and an alternative matrix. > result = rating(category_pcm, alt_mat, simulation = 500) > # rated alternatives > RM(result) scale_category idealised_priorities BSG 682 "Single" "1" CD "Single" "1" CD "Single" "1" CA9 768 "Single" "1" MUC "Single" "1" CD9 928 "Multiple" "0.5" EGFR 1956 "Single" "1" CD "Single" "1" MUC "Single" "1" FZD "Multiple" "0.5" The matrix of idealised priorities (rated alternatives) can be used to call pipeline function to estimate final priorities of alternatives. 14

15 > mat <- matrix(nrow = 7, ncol = 3, data = NA) > mat[,1] <- c('0', '1','2','3','4','4.1','4.2') > mat[,2] <- c('prioritization_of_de_genes','tumor_expression','normal_expression', + 'Frequency', 'Epitopes', 'Number_of_epitopes', 'Size_of_epitopes') > mat[,3] <- c(system.file('extdata','aggreg.judgement.tsv',package = 'Prize'), + system.file('extdata','tumor_exp_rating.tsv',package = 'Prize'), + system.file('extdata','normal_exp_rating.tsv',package = 'Prize'), + system.file('extdata','freq_exp_rating.tsv',package = 'Prize'), + system.file('extdata','epitope.pcm.tsv',package = 'Prize'), + system.file('extdata','epitope_num_rating.tsv',package = 'Prize'), + system.file('extdata','epitope_size_rating.tsv',package = 'Prize')) > # Computing alternatives priorities > prioritization <- pipeline(mat, model = 'rating', simulation = 500) References [1] T.L. Saaty. A scaling method for priorities in hierarchical structures. Journal of Mathematical Psychology, 15(3): , [2] T.L. Saaty. Rank from comparisons and from ratings in the analytic hierarchy/network processes. European Journal of Operational Research, 168(2): , January [3] T.L. Saaty. The Analytic Hierarchy Process, Planning, Piority Setting, Resource Allocation. McGraw-Hill, New york, [4] E. Forman and K. Peniwati. Aggregating individual judgments and priorities with the analytic hierarchy process. European Journal of Operational Research, 108(1): , [5] J.C. Gower. Some distance properties of latent root and vector methods used in multivariate analysis. Biometrika, 53(3/4): , [6] M. Felder, A. Kapur, J. Gonzalez-Bosquet, S. Horibata, J. Heintz, R. Albrecht, L. Fass, J. Kaur, K. Hu, H. Shojaei, R. J. Whelan, and M. S. Patankar. MUC16 (CA125): tumor biomarker to cancer therapy, a work in progress. Mol. Cancer, 13:129, [7] P. C. McDonald, J. Y. Winum, C. T. Supuran, and S. Dedhar. Recent developments in targeting carbonic anhydrase IX for cancer therapeutics. Oncotarget, 3(1):84 97, Jan